Cline
Configure Cline against RouteMux — which provider type to pick depends on the model family.
Cline can reach RouteMux two ways, and which one you need depends on the model, because RouteMux exposes each model only on the protocols it actually supports.
Which provider type
| Models you want | Cline provider type | Base URL |
|---|---|---|
anthropic/claude-* | Anthropic + Use custom base URL | https://api.routemux.com — no /v1 |
openai/*, deepseek/*, minimax/minimax-m*, google/gemini-* | OpenAI Compatible | https://api.routemux.com/v1 |
Claude models do not work through OpenAI Compatible
Every published anthropic/claude-* model is exposed on the Anthropic Messages protocol
only. Selecting Cline's OpenAI Compatible provider and typing a Claude slug returns
400 MODEL_PROTOCOL_UNSUPPORTED — the key and the base URL are
both fine, the model simply is not served on that endpoint family. This is the most common
misconfiguration we see.
Anthropic provider (for Claude models)
- Gear icon → API Provider →
Anthropic - Anthropic API Key → your RouteMux key (
sk-...) - Check Use custom base URL and enter
https://api.routemux.com - Pick a model from the dropdown
The base URL here is the bare origin: the Anthropic client appends /v1/messages itself.
Adding /v1 yourself produces /v1/v1/messages, which RouteMux does route to the right
place — but it is not what the field expects.
Cline's Anthropic model dropdown is a fixed list of Anthropic's own names. RouteMux resolves
those dialects: dashed and dotted forms are interchangeable (claude-sonnet-4-6 and
claude-sonnet-4.6), a bare name and an anthropic/-prefixed one are equivalent, and a
trailing -YYYYMMDD date snapshot is ignored. What is not forgiving: a Claude version we
do not carry will still fail. GET /v1/models is the authoritative list.
OpenAI Compatible provider (for everything else)
| Field | Value |
|---|---|
| API Provider | OpenAI Compatible |
| Base URL | https://api.routemux.com/v1 |
| API Key | your RouteMux key (sk-...) |
| Model ID | a slug from GET /v1/models, e.g. openai/gpt-5.5 |
Note that google/gemini-* text models are served on the OpenAI Chat protocol here, so they
belong in this provider, not in Cline's Gemini provider.
Long tasks and streaming
Cline runs multi-step agent loops with long streaming responses. Two facts matter:
- No total time limit on a streaming request. A turn that legitimately runs for minutes is fine.
- There is an idle watchdog: 180 s for text. If the upstream produces no new chunk for 180 seconds, RouteMux aborts the request and records it as a timeout. This is ours, on top of whatever the upstream does — a long silent think can hit it, a long talking turn cannot.
SSE is forwarded without buffering.
Troubleshooting
| Symptom | Cause |
|---|---|
400 MODEL_PROTOCOL_UNSUPPORTED | A Claude model on the OpenAI Compatible provider. Switch to the Anthropic provider — see above. |
Instant 401 | Key not saved, or copied with surrounding whitespace. |
400 AUTH_AMBIGUOUS_API_KEY | Two different key values in two header fields. Clear the one you are not using. |
422 BILLING_INSUFFICIENT_CREDITS mid-task | Wallet ran out during the run. Top up; completed steps were billed, the failed one was not. |
| Model rejected | Not enabled for your key — check GET /v1/models. |
| Task dies after a long pause with no output | The 180 s idle watchdog. Retry; if it repeats on the same prompt, the upstream is stalling rather than thinking. |